International Audio Laboratories Erlangen

Room Impulse Response Classification

Problem

Simulated room impulse responses (RIRs), most often generated with the image source method (ISM), are widely used in audio signal processing as scalable alternatives to resource-intensive measurements. Yet a systematic acoustic and perceptual gap persists between simulated and measured RIRs: geometric simulators neglect diffraction, scattering, and device-specific influences, and downstream systems trained on synthetic reverberation show a simulation-to-real gap. Moreover, no automated tool exists to distinguish these RIRs, which would be useful both for benchmarking simulated RIRs and for selecting high-quality simulations for dataset generation.

Approach

We formulate the distinction as a binary classification task and introduce RIR-SoM (RIR Simulated-or-Measured), a compact one-dimensional residual network that operates directly on RIR waveforms without handcrafted acoustic features. It consists of a ResNet encoder, a projection to a normalised embedding, and a small classification head.

  • Trained on measured and ISM-generated RIRs and evaluated under a room-held-out protocol on unseen rooms and unseen simulation methods: different ISM reflection orders, hybrid simulator using Treble SDK, and device response augmentation.
  • Separates ISM-simulated from measured RIRs almost perfectly, with progressively lower separability for the other held-out datasets.
  • Embedding analysis shows that simulated and measured datasets occupy distinct regions of the learned representation, with device-response augmented ISM positioned closest to the measured distribution.
  • Score-CAM saliency analysis (a one-dimensional adaptation) localizes the temporal regions associated with measured-label decisions, for instance the direct-path region.

The classifier can support the benchmarking of simulated RIRs and the selection of realistic simulations for dataset generation and downstream audio processing tasks.

Publications

  • Z. Xu, A. Brendel, A. G. Prinn, and E. A. P. Habets, "RIR-SoM: Probing the gap between simulated and measured room impulse responses," IWAENC 2026.